1.

Introduction

D.A. Wolf-Gladrow: LNM 1725, pp. 1–13, 2000.

c Springer-Verlag Berlin Heidelberg 2000
2 1. Introduction

1.1 Preface

Lattice-gas cellular automata (LGCA)1 and even more lattice Boltzmann

models (LBM) are relatively new and promising methods for the numerical
solution of (nonlinear) partial differential equations. Each month several pa-
pers appear with new models, investigations of known models or methodically
interesting applications. The field of lattice-gas cellular automata started al-
most out of the blue in 1986 with the by now famous paper of Frisch, Has-
slacher and Pomeau. These authors showed, that a kind of billiard game2
with collisions that conserve mass and momentum, in the macroscopic limit
leads to the Navier-Stokes equation when the underlying lattice possesses a
sufficient (hexagonal in two dimensions) symmetry. A few years later lattice
Boltzmann models arose as an offspring of LGCA. Their higher flexibility
compared to LGCA led to artificial microscopic models for several nonlinear
partial differential equations including the Navier-Stokes equation.
I have followed the exciting development of both methods since 1989 and
from time to time have given courses on this topic at the Department of
Physics and Electrical Engineering at the University of Bremen (Germany).
The present book is an extended version of my lecture manuscript.
The word ‘introduction’ in the title implies two things. Firstly, the level of
presentation should be appropriate for undergraduate students. Thus meth-
ods like the Chapman-Enskog expansion or the maximum entropy principle
which are usually not taught in standard courses in physics or mathematics
are discussed in some detail. Secondly, in an introduction many things have
to be left out. This concerns, for instance, models with several colors which
allow the simulation of multiphase flows3 or magnetohydrodynamics. Only a
few applications of LGCA or LBM to physical problems can be considered.
Interesting topics like the divergence of transport coefficients in 2D are not
discussed. The interested reader will find, however, references pointing to
original articles (especially in the ‘What else?’ sections).
The lattice-gas cellular automata require special programming techniques
which are only sparsely discussed in the widely scattered literature. The book
will hopefully fill a gap in this respect (see Subsections 3.1.2, 3.1.4, 3.2.2).
Several program codes will be made available via internet (http://www.awi-
bremerhaven.de/Modelling/LGCA+LBM/index.html).
Many mathematical text books and courses contain lots of definitions, theo-
rems and proofs - and not much else. In this respect the current book is rather
‘unplugged’: the emphasis is more focused on presenting the main principles

1
The abbreviations are explained in Table 6.6.5 on page 270.
2
Goldenfeld and Kadanoff (1999) compare the time development of lattice-gas
cellular automata with a square dance.
3
When I had almost finished my manuscript I became aware of the wonderful
book by Rothman and Zaleski (1997). Simulation of multiphase flows is a major
topic in that book.
 1.1 Preface 3

and not on teaching proof techniques. Nonetheless the proofs of several es-
sential theorems are presented in detail.
Last but not least, I would like to add a few comments on the exercises.
Problems with one star (∗) should be very easy to solve (in a few minutes);
those with two stars require more thinking or somewhat lengthy (‘... after
some algebra ...’) calculations. Exercises with three stars are very different.
Some of them require quite a bit of programming; others address more ad-
vanced stuff which has not been treated here. And finally, some of the three
star exercises point to open problems which I have not solved myself.
Acknowledgements:
The following people supported me in one or the other way by teaching me
mathematics and physics, introducing me to LGCA, providing PhD posi-
tions, asking stimulating questions, proofreading etc. Ernst Augstein, Uwe
Dobrindt, Fritz Dröge, Lars-Peer Finke, Silvia Gladrow, Vladimir Gryanik,
Wolfgang Hiller, Matthias Hofmann, Heiko Jansen, Charilaos Kougias, Ger-
rit Lohmann, Ferial Louanchi, Christof Lüpkes, Ralf Nasilowski, Dirk Olbers,
Christoph Völker, Armin Vogeler, Werner Wrede, Richard Zeebe.
I am grateful to them all.
Three anonymous referees made useful comments.
I also thank Stefanie Zöller (Springer Verlag) for support.
4 1. Introduction

1.2 Overview

The plan of the book is as follows (compage Fig. 1.2.1). In an introductory

section the Navier-Stokes equation and several approaches to solve it are dis-
cussed. In Chapter 2 cellular automata (CA) are treated in some detail in
order to show the special character of lattice-gas cellular automata. CA rules
are usually not restricted by conservation laws which is a nice feature when
simulating growth processes. The spatial propagation of properties is part of
the local updating rule. In contrast, lattice-gas cellular automata obey cer-
tain conservation laws and the updating is splitted into a local ‘collision’ and
a propagation to the nearest-neighbor sites. This splitting makes it easier to
construct models with desired macroscopic properties. The CA chapter can
be skipped in first reading.
Chapter 3 on lattice-gas cellular automata starts with the historically first
LGCA, namely the HPP model. This is the simplest model that aimed to sim-
ulate the Navier-Stokes equation (but failed to do so!). The emphasis here is
on a discussion without digging too much into theory. Special programming
techniques like multi-spin coding are explained in detail.
The FHP model is the first successful LGCA. Starting from the Boolean
microdynamics the macroscopic equations will be derived up to first order
(Euler equation) by a multi-scale expansion (Chapman-Enskog). The second
order which yields the Navier-Stokes equation will be addressed later on in
the chapters on statistical mechanics (Section 4.2) as well as in the one on
lattice Boltzmann models (Section 5.2.3).
The difference between failure (HPP) and success (FHP) depends on the
symmetry of the underlying lattice. The tensor of rank four formed from
products of the lattice vectors is part of the advection term and has to be
isotropic. The main problem in proposing a LGCA for simulations of flows in
three dimensions is to find a lattice with sufficient symmetry. In Section 3.3
the lattice tensors of rank two and four for several lattices will be calculated
and investigated for isotropy.
If one restricts oneself to single-speed models the only lattice feasible for
three-dimensional simulations is the four-dimensional face-centered hyper-
cube (FCHC). Several possible collision rules for this model are outlined in
Section 3.5. As an alternative to FCHC multi-speed models are available.
When the collision rules are carefully chosen these models conserve energy
in addition to mass and momentum and therefore are called thermal models
(Section 3.7). Another alternative for simulation in 3D is the pair-interaction
(PI) model (Section 3.6). The collision rules of this model are simple in 2D
as well as in 3D and thus allow coding using bit-operators.
In Chapter 4 some relevant concepts from statistical mechanics are discussed.
Specifically the Boltzmann equation, its five collision invariants, and its
(global) equilibrium distribution (Maxwell-Boltzmann) are presented. The
chapter contains a proof of Boltzmann’s famous H-theorem. For many ap-
 1.2 Overview 5

plications the complicated collision integral can be substituted by a relax-

ation toward equilibrium by a term that is proportional to the deviation
of the actual distribution from its (local) equilibrium. With this so-called
BGK approximation it is possible to derive the Navier-Stokes equation by
the Chapman-Enskog expansion on few pages (Section 4.2). In addition, this
chapter contains a section on the maximum entropy principle which will be
applied later on in the derivation of equilibrium distributions for lattice Boltz-
mann models.
Chapter 5 is devoted to lattice Boltzmann models. This chapter is almost
selfcontained. Readers who are only interested in LBMs (and not in LGCA)
can start here but should read at some point Section 3.3 on lattice tensors.
However, some remarks in this chapter only make sense to those who are
familiar with LGCA.
In Section 5.1 some problems with LGCA are listed and the transition from
LGCA to LBM is sketched. The section on the D2Q9 model is in some re-
spect the pendant to the FHP4 section in Chapter 3 in that this BGK model
is discussed in full detail. The equilibrium distributions are calculated from
the maximum entropy principle, the Navier-Stokes equation is derived by
Chapman-Enskog expansion and implementations of various boundary con-
ditions are discussed. This model is applied to ocean circulation in Section
5.7. The stability of the D2Q9 and other LBMs is discussed in Section 5.6.
Although the use of the maximum entropy principle is a very elegant method,
it hides the much wider freedom in choosing equilibrium distributions. Al-
ternatively, one may start from a reasonable ansatz for the distributions and
then fix the free parameters during or after the multi-scale expansion such
that the desired equations (Navier-Stokes or other partial differential equa-
tions) are obtained (Section 5.4).
This ansatz method is used to derive LBMs for diffusion equations (linear as
well as nonlinear in any number of dimensions) in Section 5.8. These models
can easily be extended to LBMs for reaction-diffusion equations. With the
same methods thermal LBMs can be constructed (Section 5.5). LBMs for
simulation in 3D are described in Section 5.3. The appendix contains a sec-
tion on Boolean algebra, some lengthy calculations and code listings of FHP
collision rules.

4
Although the underlying lattices are different!
6 1. Introduction

Fig. 1.2.1. Overview

INTRODUCTION
Chapter 1

HYDRODYNAMICS
Section 1.3

CA LGCA STATISTICAL LBM

Chapter 2 Chapter 3 MECHANICS Chapter 5
Chapter 4

BOOLEAN BOLTZMANN
ALGEBRA HPP EQUATION LGCA → LBM
Section 3.1 Section 5.1
Section 6.1 Section 4.1

MULTISPIN
CODING FHP H-THEOREM
Section 3.2 Section 4.1.2
Sections 3.1, 6.3

LATTICE BGK LBGK

TENSORS APPROX. Section 5.2
Section 3.3 Section 4.1.3

FCHC CHAPMAN- 3D-LBM

Section 3.5 ENSKOG Section 5.3
Section 4.2

MAXIMUM ANSATZ
PI ENTROPY METHOD
Section 3.6 PRINCIPLE Section 5.4
Section 4.3

THERMAL THERMAL
LGCA LBM
Section 3.7 Section 5.5

OUTLOOK
Section 5.10
1.3 The basic idea of lattice-gas cellular automata and lattice Boltzmann models 7

1.3 The basic idea of lattice-gas cellular automata and

lattice Boltzmann models

Lattice-gas cellular automata (LGCA) and lattice Boltzmann models (LBMs)

are methods for the simulation of fluid flows5 which are quite distinctive from
molecular dynamics (MD) on the one hand and methods based on the dis-
cretization of partial differential equations (finite differences, finite volumes,
finite elements, spectral methods) on the other hand. Here the basic idea
of LGCA and LBM will be sketched and the differences compared to other
methods will be outlined.

1.3.1 The Navier-Stokes equation

The flow of incompressible fluids can be described by the Navier-Stokes

equation6
∂u
+ (u∇)u = −∇P + ν∇2 u (1.3.1)
∂t
together with the continuity equation7

∇·u=0 (1.3.2)

where ∇ is the nabla operator, u is the flow velocity, P = p/ρ0 the kinematic
pressure, p the pressure, ρ0 the constant mass density and ν the kinematic
shear viscosity. Different fluids like air, water or olive oil are characterized
by their specific values of mass density and viscosity (νair = 1.5 · 10−5 m2
s−1 , νwater = 10−6 m2 s−1 , νolive oil = 10−4 m2 s−1 ). Incompressible flows
of these fluids obey the same form of equation (Navier-Stokes) whereas their
microscopic interactions are quite different (compare gases and liquids!). The
Navier-Stokes equation is nonlinear in the velocity u which prohibits its an-
alytical solution except for a few cases. Numerical methods are required to
5
... and several other processes which can be described on the macroscopic level
by partial differential equations ...
6
The viscous term of the equation was derived in different ways by Claude
Louis M. H. Navier (1785-1836) and Sir George Gabriel Stokes (1819-1903). The
Navier-Stokes equation in tensor notation reads:

∂t uα + uβ ∂xβ uα = −∂xα P + ν∂xβ ∂xβ uα .

7
∇ · u = 0 is derived from the general continuity equation
∂ρ
+ ∇ · (ρu) = 0
∂t
by setting ρ = constant.
8 1. Introduction

simulate the time evolution of flows. On the other hand, the nonlinear ad-
vection term is most welcome because it is responsible for many interesting
phenomena such as solitons (nonlinear waves), von Karman vortex streets
(regular vortex shedding behind an obstacle) or turbulence.

The Reynolds number and dynamic similarity of flows. Flows with

small velocities are smooth and are called laminar. At very high velocities
they become turbulent. The transition from laminar to turbulent flows does
not depend only on velocity as will be shown below. Consider the flow past an
obstacle, such as a sphere, a cylinder or a plate. What are the characteristic
scales of the flow? Obviously the flow field will depend on the (unperturbed)
upstream speed U and the linear size (diameter L) of the obstacle. The fluid
is characterized by its kinematic viscosity ν. The three parameters U , L and
ν have dimensions [length time−1 ], [length] and [length2 time−1 ]. It is easy
to see that from these parameters one can form essentially one dimensionless
number, namely the Reynolds number
UL
Re = . (1.3.3)
ν
The parameters U and L can be used to scale all quantities in the Navier-
Stokes equation (the primed quantities are measured in units of U and L):
u
= u/U , x
= x/L, ∇
= L · ∇, ∇
2 = L2 · ∇2 , t
= t · U/L (the advection
time scale L/U is the time for the unperturbed flow to pass the linear size of
the obstacle), P
= P/U 2 (the kinematic pressure has the dimension of energy
per mass). Inserting the scaled quanties into the Navier-Stokes equation leads
to
∂u
U 2



U 2 U2 U


+ u∇ u = −∇
P
+ ν∇
2 u
2
∂t L L L L
or after division by U 2 /L
∂u




1
2



+ u ∇ u = −∇
P
+ ∇ u. (1.3.4)
∂t Re
The scaled Navier-Stokes equation equation (1.3.4) does not contain any scale
and only one dimensionless quantity, namely the Reynolds number. Thus for
a given type of flow (say the flow past a sphere) the scaled velocity of a
stationary flow will depend only on the scaled spatial coordinate and the
Reynolds number: x 
u
u
= = fu , Re (1.3.5)
U L
where the function fu depends on the geometry of the problem (the type of
flow). The same is true for the scaled pressure:
P x 
P
= 2 = fP , Re . (1.3.6)
U L
1.3 The basic idea of lattice-gas cellular automata and lattice Boltzmann models 9

Thus all flows of the same type but with different values of U , L and ν are
described by one and the same non-dimensional solution (u
, P
) if their
Reynolds numbers are equal. All such flows are said to be dynamically simi-
lar.
The value of the Reynolds number provides an estimate of the relative im-
portance of the non-viscous and viscous forces. The pressure gradient usually
plays a passive role, being set up in the flow as a consequence of motions of
a rigid boundary or of the existence of frictional stresses (Batchelor, 1967).
Thus the flows can be characterized by the relative magnitudes of advection
and viscous forces:
|(u∇) u| U 2 /L U ·L
≈ = = Re . (1.3.7)
|ν∇2 u| νU/L2 ν
Flows with small Reynolds numbers (Re  1) are laminar, von Karman
vortex streets are observed at intermediate values (Re ≈ 100) and turbulent
flows occur at very high Reynold numbers (Re  1). The fact that flows can
be characterized by Re and the law of dynamic similarity were first recognized
by Stokes (1851) and Reynolds (1883).
The law of dynamic similarity provides the link between flows in the real
world where length is measured in meters and the simulation of these flows
with lattice-gas cellular automata and lattice Boltzmann models over a lattice
with unit grid length and unit lattice speed. In these models the viscosity is
a dimensionless quantity because it is expressed in units of grid length and
lattice speed. These dimensionless flows on the lattice are similar to real flows
when their Reynolds numbers are equal.

1.3.2 The basic idea

The fact that different microscopic interactions can lead to the same form of
macroscopic equations is the starting point for the development of LGCA. In
addition to real gases or real liquids one may consider artificial micro-worlds
of particles ‘living’ on lattices with interactions that conserve mass and mo-
mentum. The microdynamics of such artificial micro-worlds should be very
simple in order to run it efficiently on a computer. Consider, for example,
a square lattice with four cells at each node such that one cell is associated
with each link to the next neighbor node (compare Fig. 3.1.1 on page 41).
These cells may be empty or occupied by at most one particle with unit mass
m = 1. Thus each cell has only two possible states and therefore is called
a cellular automaton. Velocity and thereby also momentum can be assigned
to each particle by the vector connecting the node to its next neighbor node
along the link where the particle is located. These vectors are called lattice
velocities. The microscopic interaction is strictly local in that it involves only
particles at a single node. The particles exchange momentum while conserv-
ing the mass and momentum summed up over each node. After this collision
10 1. Introduction

each particle propagates along its associated link to its next neighbor node.
The microdynamics consists on a repetition of collision and propagation.
Macroscopic values of mass and momentum density are calculated by coarse
graining (calculation of mean values over large spatial regions with hundreds
to thousands of nodes).
Do these mean values obey the Navier-Stokes equation? The answer is neg-
ative for the model just sketched (discussed in more detail in Section 3.1).
This model was proposed by Hardy, de Pazzis and Pomeau in 1973 (HPP
model). It took more than 10 years before Frisch, Hasslacher and Pomeau
(1986) found the third essential condition in addition to mass and momen-
tum conservation: the lattice has to possess a sufficient symmetry in order
to ensure isotropy of a certain tensor of fourth rank formed from the lattice
velocities. In 2D, for example, 4-fold rotational symmetry (square lattice) is
not enough whereas hexagonal symmetry (triangular lattice; FHP model; see
Section 3.2) is sufficient.
A further condition should be mentioned here. The microdynamics must not
possess more invariants than required by the desired macroscopic equations
because such so-called spurious invariants can alter the macroscopic behavior
by unphysical constraints (compare Section 3.8).
The importance of the work of Frisch, Hasslacher and Pomeau (1986) can
hardly be overestimated. Their finding of the lattice symmetry condition
started an avalanche of LGCA models. Finding a lattice with sufficient sym-
metry for simulations in 3D was a tough job. Wolfram (1986) showed that
lattice tensors over the face-centered hypercube (FCHC) are isotropic up to
rank 4.
Lattice Boltzmann models were first based on LGCA in that they used the
same lattices and applied the same collisions. Instead of particles, LBMs deal
with continuous distribution functions which interact locally (only distribu-
tions at a single node are involved) and which propagate after ‘collision’ to
the next neighbor node. Coarse graining is not necessary any more. In the
beginning this was considered as the main advantage of LBMs compared to
LGCA. The next step in the development was the simplification of the col-
lision operator and the choice of different distribution functions. This gives
much more flexibility of LBMs, leads to Galilei invariant macroscopic equa-
tions without scaling of time, and allows to tune viscosity. Most recently
LBMs living on curvilinear coordinate systems have been proposed.

Exercise 1.3.1. (**)

Consider flows that are affected by an external force such as gravity. Discuss
the consequences for the similarity of flows. How many independent dimen-
sionless numbers are required to characterize the flow?
1.3 The basic idea of lattice-gas cellular automata and lattice Boltzmann models 11

1.3.3 Top-down versus bottom-up

The conventional simulation of fluid flows (and other physical processes)

generally starts from nonlinear partial differential equations (PDEs). These
PDEs are discretized by finite differences (Ames, 1977; Morton and May-
ers, 1994), finite volumes (Bryan, 1969), finite elements8 (Zienkiewicz, and
Taylor, 1989 and 1991), or spectral methods (Machenhauer, 1979; Bourke,
1988). The resulting algebraic equations or systems of ordinary differential
equations are solved by standard numerical methods. Although this ‘top-
down’ approach seems to be straightforward it is not without difficulties. In
many textbooks on the numerical solution of partial differential equations
the authors put much emphasis on the truncation error which is due to the
truncation of Taylor series when going from differential to finite differences
whereas physicists are usually more concerned whether or not certain quan-
tities are conserved also by the discretized form of the equations. This latter
property is most important for integrations over long time scales in closed
domains like, for instance, in the simulation of the world oceans or in coupled
atmosphere-ocean models. A small leakage would transform the ocean into
an empty basin after some time. Numerical instabilities are another prob-
lem of this type of numerical methods (Courant, Friedrichs and Lewy, 1928;
Phillips, 1956 and 1959).
LGCA and LBM are different variants of the ’bottom-up’ approach (Fig.
1.3.1) where the starting point is a discrete microscopic model which by con-
struction conserves the desired quantities (mass and momentum for Navier-
Stokes equation). These models are unconditional stable (LGCA) or show
good stability properties (LBM). The derivation of the corresponding macro-
scopic equations requires, however, lengthy calculations (multi-scale analy-
sis). A major problem with the bottom-up approach is to detect and avoid
spurious invariants which is, by the way, also a problem for the models de-
rived by the top-down approach. The construction of LGCA or LBM for
given macroscopic equations seems to require some intuition. Meanwhile at
least for LBM there exists a recipe for the construction of appropriate micro-
dynamics when the conserved quantities of the physical process are known
(compare Section 5.4).

1.3.4 LGCA versus molecular dynamics

Another bottom-up approach is molecular dynamics (MD) (Verlet, 1967;

Evans and Morriss, 1983; Heyes et al., 1985; Mareschal and Kestemont, 1987;
Boon and Yip, 1991; Rapaport, 1995). In MD one tries to simulate macro-
scopic behavior of real fluids by setting up a model which describes the mi-
croscopic interactions as good as possible. This leads to realistic equations of
8
These finite methods can be combined with multigrid techniques; see, for exam-
ple, Hackbusch (1985).
12 1. Introduction

state whereas LGCA or LBM posses only isothermal relations between mass
density and pressure. The complexity of the interactions in MD restricts the
number of particles and the time of integration. A method somewhat in be-
tween MD and LGCA is maximally discretized molecular dynamics proposed
by Colvin, Ladd and Alder (1988).
1.3 The basic idea of lattice-gas cellular automata and lattice Boltzmann models 13

Fig. 1.3.1. Top-down versus bottom-up (see text).

PARTIAL PARTIAL
DIFFERENTIAL DIFFERENTIAL
EQUATIONS EQUATIONS
(NAVIER-STOKES) (NAVIER-STOKES)

DISCRETIZATION MULTI-SCALE ANALYSIS

DIFFERENCE
EQUATIONS DISCRETE MODEL
(CONSERVED (LGCA OR LBM)
QUANTITIES?)